Stanford researchers slash robot training data needs by 70%

Researchers introduced Counterfactual Nuisance Behaviour Cloning (CFNBC), an offline data-selection framework for targeted robustness repair in robot imitation learning. Instead of simply adding more demos, CFNBC generates paired clean and nuisance observations that preserve the expert action, then measures "action drift"—how much the policy's predicted action changes under a nuisance that shouldn't affect the behavior. This pinpoints policy-specific fragile response modes without needing rollouts or success labels. In MuJoCo bimanual cube transfer and SimplerEnv cube stacking, action drift predicted nuisance-induced failures, and selecting just 20–30 targeted candidates outperformed random selection at the same budget—approaching the performance of much larger random repair sets.